This paper presents a method to extract causal relationships of events from Twitter. We extracted event-speci c words, which are frequently used in a speci c period, from tweet archives. Next, we make a series of event-speci c words for each user and make a transition relationship matrix by counting their anteroposterior relationships between event-speci c words. Existence or nonexistence of causality, its direction, and its strength are determined by analyzing a transition relationship matrix. Furthermore, we simplify an extracted graph structure by removing redundant causal edges. In fact, we make a causal relationship network from tweet archive in the Great East Japan Earthquake. We analyze the network structure and show that proposed method is suitable for extracting causal relationships.